Tools

Apple's own array framework, built so unified memory stops being a compromise.

The package behind every `mlx_lm.generate` command on this site.

MLX's tensor core, ported so Swift apps don't need a Python bridge.

The C++ runtime that made GGUF and Metal inference normal.

A compiler-first runtime that targets Metal, CUDA and mobile from one graph.

Hugging Face's minimalist Rust tensor library, Metal backend included.

The `ollama run` command that made local models a one-liner.

An OpenAI-compatible server that's actually running MLX underneath.

LM Studio's own MLX backend, open-sourced separately from the app.

A ChatGPT-shaped front end you point at any local backend.

Turns a pile of Macs (and phones) into one inference cluster.

An open-source, offline-first ChatGPT alternative with its own engine.

The kitchen-sink UI power users reach for when they want every knob.

The same mlx-lm package doubles as Apple's official fine-tuning path.

The fine-tuning library everyone quotes for faster training on less memory.

The YAML-config trainer that treats every method as a config file.

Hugging Face's parameter-efficient fine-tuning library, the one everything wraps.

One command turns a bf16 Hugging Face repo into a quantized MLX folder.

Apple's converter and compressor for getting models onto the Neural Engine.

Model-merging toolkit that treats checkpoints like Lego bricks.

Text-to-speech, speech-to-text and audio models under one MLX package.

A batched, quantized Whisper reimplementation chasing raw transcription speed.

Prompt-controllable speech: describe the voice you want in plain English.

Speech in, speech out, skipping the ASR-LLM-TTS relay race.

The go-to package for running vision-language models with MLX.

Apple's on-device-sized CLIP for fast image-text matching.

The reproducible, community-trained CLIP zoo everyone benchmarks against.

Apple's research VLM for referring and grounding inside images.

Argmax's toolkit for running diffusion models fast via MLX and Core ML.

Apple's own Core ML conversion pipeline for Stable Diffusion.

The reference SwiftUI app that proved diffusion runs natively on Mac.

Embedding models on MLX, so retrieval doesn't need PyTorch.

An all-in-one embeddings database, from before 'RAG' was a word.

The data framework most local RAG tutorials are built on.

The vector database that got embedded into every local RAG demo first.

Working Swift ports of the MLX examples: LLMs, VLMs, diffusion.

Hugging Face's Swift package for tokenizers, hub downloads and generation.

The Swift package that made on-device Whisper feel like an Apple API.

Microsoft's multi-agent conversation framework, one of the first to stick.

LangChain's answer to 'agents need state machines, not chains.'

Role-based orchestration: give each agent a job title and a goal.

The agent scaffold built specifically to fix real GitHub issues.